A fly's wiring diagram, run untrained, already knows how to behave

What the work claims

This is a primary computational result from John Minnick, Jesus Gonzalez-Ferrer, Jinghui Geng, Jason Eshraghian, Mohammed Mostajo-Radji, David Haussler, and Mircea Teodorescu. The claim is deliberately conservative in method and sweeping in implication: run a connectome exactly as read, with zero fitting, and it reproduces a catalog of known biology, supports whole-body navigation in a physics simulation, and even shows a basic form of associative learning when one class of synapse is allowed to weaken under a dopamine-like gate.¹

Outside the body, the group drove 15 pathways whose behavioral role is known from experiments in the real fly, under criteria written down before each test. Seven conducted, died under a wiring shuffle, and carried the sign measured in the animal; a pathway conducts when it supplies roughly 2% or more of its target neuron's input, while those supplying a hundredth of that fail.¹ The failures are as informative as the successes: the giant fibre escape pathway only works once an electrical synapse the connectome does not record is added by hand, and steering pathways need background activity a silent network cannot supply.¹

In the body (the published NeuroMechFly physics model), twenty simulated flies crossed nine arenas holding 32 odour sources behind walls. The intact network reached 30.25 of 32 sources per fly (95%); a rewired twin preserving every neuron's connection counts and signs reached 5.80 (18%); a mushroom-body-lesioned network reached 3.45; a correlated random walk reached 2.20.¹ Pinning a single scalar, the approach-avoidance balance read out of the mushroom body, rescued the lesioned and rewired flies, locating their deficit at that output.¹ Replaying intact motor commands with sensory feedback cut reached about 2% of sources, so navigation needs the closed loop, not just good commands.¹ And dopamine-gated weakening of mushroom-body output synapses flipped attraction to avoidance, associative in behavior only when the odour was present during reinforcement (paired versus reinforcement alone: p = 6.8 times 10 to the minus 8, effect size delta = +1.00 over 20 flies per group).¹

How it works

The substrate is the BANC connectome, an electron-microscopy reconstruction of one adult female fly's brain and nerve cord: 172,433 neurons, 11,752,828 connections, 35,733,096 synapses, each neuron signed excitatory or inhibitory by a published transmitter prediction.¹ Every neuron is a leaky integrate-and-fire unit whose parameters (membrane constants, thresholds, synaptic strengths) come from a previously published whole-brain fly model, and none of those values was adjusted for this study.¹ This "as read" discipline is the methodological core: the test of what wiring explains is only meaningful if nothing else was tuned to pass it.

The body coupling is two fixed rules written once and never varied: arena odours and looming objects set the firing rates of sensory neurons, and six leg motor-neuron pools set the gait, with turns scaled by a valence signal computed as the balance between mushroom-body output neurons that drive approach and those that drive avoidance.¹ That is the entire interface. The rewiring control shuffles connections while preserving each neuron's in-degree, out-degree, and the excitatory or inhibitory sign of every link, so any difference in behavior is attributable to which neurons connect, not to global statistics.¹ Learning is implemented at one site only, the mushroom body's output synapses, weakened under a dopamine gate while the fly forages, mirroring the mechanism biology established for olfactory learning in the real fly.¹

One honest wrinkle the authors surface rather than hide: read at face value, the antennal lobe's local neurons form a self-sustaining excitatory loop, and 68% of Kenyon cells fire to every odour, where the real fly's code is sparse at 5 to 10%.¹ A single scaling factor (0.15) applied to that one synapse class brings the code into the fly's measured range, and the associative learning result holds only in this matched variant; in the strictly as-read network, reinforcement without the odour depresses the same synapses almost as much, so the learning is not associative there.¹

Where a skeptic should push

The most load-bearing assumption is that a parameter set borrowed from a different model transfers to this connectome. The neuron model and its constants come from a published fly-brain model built partly on functional data; if those constants are systematically off for BANC's cell types, the "untrained" result inherits those errors, and the 2% conductance boundary is an empirical property of this particular parameterization, not a law of wiring diagrams. The authors say exactly this in the paper.¹

Second, the embodied claim has a scope the authors themselves police carefully. The body interface contains real computation: the turn is computed from the odour's left-right difference and scaled by the mushroom-body valence, and the gait is the body model's. The authors state plainly that their rescue experiments "do not show that the connectome computes the turn," only what the connectome supplies to that rule.¹ The navigation result is therefore a statement about a hybrid system: wiring plus a two-rule controller. Third, replication is across simulation seeds of a single individual's connectome, not across animals; one fly's wiring, twenty random initial conditions. Fourth, the known annotation errors are not trivial: the transmitter prediction disagrees with verified signs for 2,732 of the 60,448 neurons that have one, and the one cell whose sign is known to be mis-predicted (the GABAergic APL neuron) is directly in the olfactory pathway under study.¹ The learning result, in particular, lives entirely inside the matched variant whose single tuned factor fixes an as-read pathology; treat it as a proof of mechanism, not as untrained emergence.

Wiring, not training, may hold the compute value

The commercial bet in organoid intelligence is that living neural tissue earns its keep through plasticity: you buy the tissue, you train it on your task, the learning is the product. This paper complicates that bet in a way neither the hype nor the skeptics predicted. In the most completely mapped nervous system available, the competence to navigate, avoid obstacles, and gate approach by an internal valence signal is present before any task-specific adjustment of the network: 95% of targets with the measured wiring, 18% with statistically matched but shuffled wiring, and the deficit is a single readout scalar that the wiring feeds.¹ If a large share of useful behavior is architectural rather than learned, then the value in a neural substrate sits in its connectivity, and "we train your tissue" is a weaker sales pitch than "our tissue comes wired."

The vendor-capability consequences cut both ways. Opportunity: the rewiring control is a template for a causal quality test any platform could run. Preserve every node degree and every sign, shuffle the connections, and compare; the gap between intact and shuffled performance is the amount of behavior the specific architecture carries, as opposed to generic network statistics.¹ A dish-grown neural platform that can demonstrate such a gap has evidence that its tissue is not merely "a recurrent network," the perennial dismissal of the field; one that cannot has been handed a falsifiable self-check it previously lacked. Threat: architecture is exactly the property that is expensive to produce and easy to hoard. An electron-microscopy connectome of this completeness costs years of imaging and annotation; whoever controls the wiring diagram controls the substrate's decisive asset, and unlike weights in a software model, a connectome is not copyable from an API. Expect the platform layer of this field to stratify into wiring owners and wiring tenants.

There is also a direct challenge to the living-tissue value proposition. If a frozen, static, imperfectly annotated wiring diagram, simulated on silicon at zero marginal energy for culturing, already navigates at 95%, then "it computes" cannot be the argument for maintaining living neural tissue. The living substrate has to justify itself on properties the diagram lacks: the self-repair, metabolic autonomy, and lifelong plasticity that a fixed connectome does not have, and, on the evidence here, genuinely needs. Recall that replayed commands without sensory feedback reached almost no targets even though the commands were intact: the behavior lived in the loop, in continuous exchange between body and network, not in a stored program.¹ That is precisely the kind of capability a living, sensing, self-maintaining substrate is positioned to offer and a snapshot simulation is not. The honest pitch for biocomputing after this paper is not that neurons compute, but that living tissue stays in the loop by construction.

Finally, a governance footnote worth recording before it becomes a dispute. A connectome is one particular individual's anatomy. As neural platforms acquire patterned architecture, "whose wiring is this" will follow the same arc as cell-line provenance: a question of attribution, consent, and ownership that regulators already know how to ask about tissue and have never had to ask about computation. The fly answered it by having a diagram read like a score and played without a rehearsal. The organoid field will not get that transparency by default; it will have to build it.

The bottom line

Established, from the verified full text: an all-unfitted spiking model of a complete fly brain-and-nerve-cord connectome reproduces seven of fifteen experimentally characterized behavioral pathways at a conductance threshold of about 2% of target input; navigates a physics-simulated body to 95% of odour sources versus 18% for degree-and-sign-preserving rewiring and about 2% for replayed commands without feedback; and shows associative approach-to-avoidance reversal when a single output synapse class is weakened under a dopamine gate in a variant with one rescaled local circuit.¹ Not established: that the embodied behavior is computed by the connectome rather than by the hybrid of connectome plus two-rule controller; that parameters transfer across connectomes; or that the associative learning emerges without the one matched adjustment.¹ What would confirm the architectural-value thesis: the same protocol on a second individual's connectome, showing the 95-versus-18 gap is wiring-specific, not parameterization luck. What would break it: a demonstration that moderate parameter refitting erases the rewiring gap, making the behavior a property of the neuron model rather than the anatomy. For platform governance, the usable insight is already here and cheap: shuffle-control comparisons belong in every neural platform's reported evidence, and the gap they measure is the strongest single number the field has yet been offered for what architecture itself is worth.